source: https://x.com/aaron_defazio/status/2084779496549548323 captured: 2026-08-18T07:36:59.980126+00:00 capture: xdk genre: conceptual-essay type: kb/sources/types/snapshot.md tags: [x-article] status_id: 2084779496549548323 conversation_id: 2084779496549548323 post_count: 1
A new Era Of Theory-Driven AI Research
Author: @aaron_defazio Post: https://x.com/aaron_defazio/status/2084779496549548323 Created: 2026-08-04T23:12:19.000Z
TLDR: Empirical Explorative Research has always led the charge, with theory that has explanatory power lagging years behind. This is rapidly changing. We have now crossed the threshold where developing theory has become orders of magnitude faster. It is now possible to rapidly develop theory that can drive AI research, not just explain it post-hoc. Theory work in AI research has fundamentally changed. Past tense. The tipping point was this month. Multiple frontier models are now capable of doing useful mathematics beyond toy problems. They are perhaps not as smart as the very best researchers, but they are FASTER, and they directly automate much of the grunt-work that previously would have taken days or weeks. I have seen this first hand. I have a number of theory problems that I have been thinking on-and-off about for the last few years. This week I fed all of them into a coding agent, and in less than an hour, for every single problem some sort of major progress resulted. These are not Riemann-Hypothesis hard, but they are problems that require weeks of dedicated research effort to make progress on. It’s easy to be dismissive of the impact of this sort of progress on mathematics by AI. The recently announced progress on finding counter-examples to major open conjectures is remarkable, but like any Pure Math research it seems far removed from any immediate application. However, the progress here is not limited to pure math, that’s just where it’s most visible. These same tools are directly useful for Applied Math where the impact is about to be felt far more concretely. Theory with Predictive Power Changes Everything The goal of my research has always been to develop theory that can predict the outcome of experiments before they are run. We should be able to predict loss values, gradient norms, weight norms, and other relevant quantities. Ultimately, the theory should guide the choice of all relevant hyper-parameters, removing guess work and parameter sweeps. Theory should even be able to guide the development of new training algorithms, new architectures and new losses. It seems kind of weird to state this, because, after all, isn’t this the goal of all theory work in an applied field like AI Research? Shouldn’t theory guide experimentation? Historically, it hasn’t been the case. Most applied theory work has at best been Explanatory, more like the work of historians trying to make sense of what happened. We take an approach that has recently been adopted by practitioners and found effective. We develop a theory that explains why this approach works. This sort of post-hoc explanation unfortunately is often equally able to explain why many other approaches work, even those that, well, don’t! Developing theory with predictive power is exceptionally difficult. There are only a few successful examples in our field. It requires exploration, failure after failure after failure and generally more open-ended work than simply analyzing the behavior of an existing method in a known setting. It takes TIME. Lots of time. Months or years of work. Exactly the kind of grunt work that AI agents have recently become capable of. As the time cost of producing predictive theory comes down, we will soon reach the point where theory can start leading rather than lagging practice. This is when the impact of improvements to machine intelligence start to meaningfully accelerate. When the AI can actually understand why and how every aspect of its training pipeline contributes to its performance. Not from fragile scaling-law fits, but from theory that captures the interplay between architecture, data and optimizer. Auto-Research requires predictive theory There has been a wave of recent auto-research efforts focused on having AI Agents act in a loop, optimizing the performance of a benchmark problem through an iterative process of experimentation. The agent codes up an idea, runs it, waits on the results, then repeats. This sort of iterative experimentation is bottlenecked by the time required for the experiments to run. This bottleneck is only going to get more severe. The rapid improvements in intelligence per watt are far outpacing the (Moore’s-law bound) training speedups. Soon, running a single small experiment will feel (from the point of view of the AI agent) like building the Large Hadron Collider. A monumental undertaking. It will be worth taking the time (equivalent, for the agent, to many human lifetimes of thinking) to understand better the nature and expected results of an experiment before running it. Making good use of this thinking time is only possible if predictive theory can be developed. Otherwise, it’s just a process of informed guess-work, a random-walk descent. The Role of the Researcher With so much of the job automated already, what is the role of the researcher? What is there left to do, when AI can already run experiments and prove theorems? I would argue that the job has not changed. If anything, the automation only makes it more apparent what tasks are worth spending time on and what are not. Many folks online are worrying about how NeurIPS and ICML are about to be flooded by incremental papers that few will read, and that don’t contribute meaningfully to the advancement of the field. They talk as if NeurIPS wasn’t already flooded with such papers! The ability to produce larger volumes of incremental research won’t make the impactful research any less apparent. I often see junior researchers spend months on technical theory development of ideas just because the details are interesting. A theory doesn’t yet exist, but could. The path forward is clear. Only at the end do they test the ideas out with a few experiments, trying to jam reality through a theory-shaped hole. This work was never a good use of time, and we are already at the point where AI can essentially automate that work. Instead, more time needs to be spent on brainstorming and planning, forming specific quantifiable research goals, informed by experiment. The best researchers have already been operating in this mode. Allocating more of their time to idea exploration, brainstorming. Not following a simple theory->experiment pipeline but a rapid back-and-forth where ideas are only fully explored when they show promise. For those researchers, AI agents just accelerate their existing workflow, not fundamentally change it. The role of the researcher is clearer than ever.